-
Notifications
You must be signed in to change notification settings - Fork 1
2018 01 29 webex
Dan presented a plenary on the Sessions -> endpoints ideas. The Forum decided that the two concepts are, and should be, separate. The key difference is in the scope of isolation. Different endpoints must communicate, whereas different sessions must not communicate or interact.
Howard went back through old minutes and pointed out action items including support for legacy
apps and MPI_COMM_CREATE_TEAM in a multi-threaded environment. Dan reminds us that the
legacy support issue arose from confusion about the example. For the later point, the
MPI_COMM_CREATE_TEAM retains the possibility of multiple endpoints being supplied. Dan
will pull in some of the material from Jeff's old slides that still are applicable, including
support for legacy applications, etc.
The WG went back to the most recent version of Jeff's long slide-deck. We went through from the beginning, looking for outstanding issues and tidying up the notes. We got to slide 48 (of 105).
- Adopt the term TEAM instead of "set of processes"? Pro: proper name; Con: confusable with OpenMP [throughout]
- Thread-level per session - may require multiple MPI libraries (or MPI binaries) [slide 2]
- Which objects are global, and which just have access to globally-significant information? [slide 19]
- Can we (should we) support getting named sets for which the local process is not a member? [slide 30]